* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download Parameter Estimation with Expected and Residual-at
Uncertainty principle wikipedia , lookup
Computational electromagnetics wikipedia , lookup
Birthday problem wikipedia , lookup
Exact cover wikipedia , lookup
Genetic algorithm wikipedia , lookup
Knapsack problem wikipedia , lookup
Lateral computing wikipedia , lookup
Drift plus penalty wikipedia , lookup
Perturbation theory wikipedia , lookup
Computational complexity theory wikipedia , lookup
Generalized linear model wikipedia , lookup
Multi-objective optimization wikipedia , lookup
Travelling salesman problem wikipedia , lookup
Inverse problem wikipedia , lookup
Weber problem wikipedia , lookup
To appear in: Systems and Control Letters (Elsevier) Parameter Estimation with Expected and Residual-at-Risk Criteria Giuseppe Calafiore∗, Ufuk Topcu†, and Laurent El Ghaoui‡ Abstract In this paper we study a class of uncertain linear estimation problems in which the data are affected by random uncertainty. In this setting, we consider two estimation criteria, one based on minimization of the expected !1 or !2 norm residual and one based on minimization of the level within which the !1 or !2 norm residual is guaranteed to lie with an a-priori fixed probability (residual at risk). The random uncertainty affecting the data is characterized by means of its first two statistical moments, and the above criteria are intended in a worst-case probabilistic sense, that is worst-case expectations and probabilities over all possible distribution having the specified moments are considered. The ensuing estimation problems can be solved efficiently via convex programming, yielding exact solutions in the !2 norm case and upper-bounds on the optimal solutions in the !1 case. Keywords: Uncertain least-squares, random uncertainty, robust convex optimization, value at risk, !1 norm approximation. 1 Introduction To introduce the problem treated in this paper, let us first consider a standard parameter estimation problem where an unknown parameter θ ∈ Rn is to be determined so to minimize a norm residual of the form #Aθ −b#p , where A ∈ Rm,n is a given regression matrix, b ∈ Rm is a measurement vector, and # ·# p denotes the "p norm. In this setting, the most relevant and widely studied case arise of course for p = 2, where the problem reduces to classical least-squares. The case of p = 1 also has important applications due to its resilience to outliers and to the property of producing “sparse” solutions, see for instance [5, 7]. For p = 1, the solution to the norm minimization problem can be efficiently computed via linear programming, [3, §6.2]. In this paper we are concerned with an extension of this basic setup that arises in realistic cases where the problem data A, b are imprecisely known. Specifically, we consider the situation where the entries of A, b depend affinely on a vector δ of random uncertain . . parameters, that is A = A(δ) and b = b(δ). Due its practical significance, the parameter estimation problem in the presence of uncertainty in the data has attracted much attention G. Calafiore is with the Dipartimento di Automatica e Informatica, Politecnico di Torino, Italy. e-mail: [email protected] † U. Topcu is with the Department of Mechanical Engineering, University of California, Berkeley, CA, 94720 USA. e-mail: [email protected] ‡ Laurent El Ghaoui is with the Department of Electrical Engineering and Computer Science, University of California, Berkeley, CA, 94720, USA. e-mail: [email protected]. ∗ 1 To appear in: Systems and Control Letters (Elsevier) in the literature. When the uncertainty is modeled as unknown-but-bounded, a min-max approach is followed in [8], where the maximum over the uncertainty of the "2 norm of the residual is minimized. Relations between the min-max approach and regularization techniques are also discussed in [8] and in [12]. Generalizations of this approach to "1 and "∞ norms are proposed in [10]. In the case when the uncertainty is assumed to be random with given distribution, a classical stochastic optimization approach is often followed, whereby a θ is sought that minimizes the expectation of the "p norm of the residual with respect to the uncertainty. This formulation leads in general to numerically “hard” problem instances, that can be solved approximately by means of stochastic approximation methods, see, e.g., [4]. In the special case where the squared Euclidean norm is considered, instead, the expected value minimization problem actually reduces to a standard least-squares problem, which has a closed-form solution, see [4, 10]. In this paper we consider the uncertainty to be random and we develop our results in a “statistical ambiguity” setting in which the probability distribution of the uncertainty is only known to belong to a given family of distributions. Specifically, we consider the family of all distributions on the uncertainty having a given mean and covariance, and seek results that are guaranteed irrespective of the actual distribution within this class. We address both the "2 and "1 cases, under two different estimation criteria: the first criterion aims at minimizing the worst-case expected residual, whereas the second one is directly tailored to control residual tail probabilities. That is, for given risk $ ∈ (0, 1), we minimize the residual level such that the probability of residual falling above this level is no larger than $. The rest of the paper is organized as follows: Section 2 sets up the problem and gives some preliminary results. Section 3 is devoted to the estimation scheme with worst-case expected residual criterion, while Section 4 contains the results for the residual-at-risk criterion. Numerical examples are presented in Section 5 and conclusions are drawn in Section 6. An Appendix contains some of the technical derivations. Notation The identity matrix in Rn,n and the zero matrix in Rn,n are denoted as In and 0n , respectively (subscripts may be omitted when they can easily be inferred from context). #x#p denotes the standard "p norm of vector x; #X#F denotes the Frobenius norm of matrix √ X, that is #X#F = Tr X " X, where Tr is the trace operator. The notation δ ∼ (δ̂, D) means that δ! is a random vector with expected value E {δ} = δ̂ and covariance matrix " . " var {δ} = E (δ − δ̂)(δ − δ̂) = D. The notation X ' 0 (resp. X ( 0) indicates that matrix X is symmetric and positive definite (resp. semi-definite). 2 Problem setup and preliminaries Let A(δ) ∈ Rm,n , b(δ) ∈ Rm be such that q # . [A(δ) b(δ)] = [A0 b0 ] + δi [Ai bi ], (1) i=1 where δ = [δ1 · · · δq ]" is a vector random uncertainties, [A0 b0 ] represents the “nominal” data, and [Ai bi ] are the matrices of coefficients for the uncertain part of the data. Let 2 To appear in: Systems and Control Letters (Elsevier) θ ∈ Rn be a parameter to be estimated, and consider the following norm residual (which is a function of both θ and δ): . fp (θ, δ) = #A(δ)θ − b(δ)#p (2) = # [(A1 θ − b1 ) · · · (Aq θ − bq )] δ + (A0 θ − b0 )#p . = #L(θ)z#p , . where we defined z = [δ " 1]" , and L(θ) ∈ Rm,q+1 is partitioned as . L(θ) = [L(δ) (θ) L(1) (θ)], (3) with . L(δ) (θ) = [(A1 θ − b1 ) · · · (Aq θ − bq )] ∈ Rm,q , . L(1) (θ) = A0 θ − b0 ∈ Rm . (4) In the following we assume that E {δ} = 0 and var {δ} = Iq . This can be done without loss of generality, since data can always be pre-processed so to comply with this assumption, as detailed in the following remark. Remark 1 (Preprocessing the data) Suppose that the uncertainty δ is such that E {δ} = δ̂ and var {δ} = D ( 0, and let D = QQ" be a full-rank factorization of D. Then, we may write δ = Qν + δ̂, with E {ν} = 0, var {ν} = Iq , and redefine the problem in terms of uncertainty ν ∼ (0, I), with L(δ) (θ) = [(A1 θ − b1 ) · · · (Aq θ − bq )]Q, L(1) (θ) = [(A1 θ − b1 ) · · · (Aq θ − bq )]δ̂ + (A0 θ − b0 ). & We next state the two estimation criteria and the ensuing problems that are tackled in this paper. Problem 1 (Worst-case expected residual minimization) Determine θ ∈ Rn that minimizes supδ∼(0,I) E {fp (θ, δ)}, that is solve minn sup E {#L(θ)z#p }, θ∈R . z " = [δ " 1], (5) δ∼(0,I) where p ∈ {1, 2}, L(θ) is given in (3), (4), and the supremum is taken with respect to all possible probability distributions having the specified moments (zero mean and unit covariance). In some applications, such as in financial Value-at-Risk (V@R) [6, 11], one is interested in guaranteeing that the residual remains “small” in “most” of the cases, that is one seeks θ such that the corresponding residual is small with high probability. An expected residual criterion such as the one considered in Problem 1 is not suitable for this purpose, since it concentrates on the average case, neglecting the tails of the residual distribution. The second criterion that we consider is hence focused on controlling the risk of having residuals above some level γ ≥ 0, where risk is expressed as the probability Prob {δ : f (θ, δ) ≥ γ}. Formally, we state the following second problem. Problem 2 (Guaranteed residual-at-risk minimization) Fix a risk level $ ∈ (0, 1). Determine θ ∈ Rn such that a residual level γ is minimized while guaranteeing that Prob {δ : fp (θ, δ) ≥ γ} ≤ $. That is, solve min γ θ∈Rn ,γ≥0 subject to : supδ∼(0,I) Prob {δ : #L(θ)z#p ≥ γ} ≤ $, 3 . z " = [δ " 1], To appear in: Systems and Control Letters (Elsevier) where p ∈ {1, 2}, L(θ) is given in (3), (4), and the supremum is taken with respect to all possible probability distributions having the specified moments (zero mean and unit covariance). A key preliminary result opening the way for the solution of Problem 1 and Problem 2 is stated in the next lemma. This lemma is a powerful consequence of convex duality, and provides a general result for computing the supremum of expectations and probabilities over all distributions possessing a given mean and covariance matrix. Lemma 1 Let S ⊆ Rn be a measurable set (not necessarily convex), and φ : Rn → R a measurable function. Let z " = [x" 1], and define . Ewc = . Pwc = sup E {φ(x)} x∼(x̂,Γ) sup Prob {x ∈ S} x∼(x̂,Γ) . Q = $ Γ + x̂x̂" x̂ x̂" 1 % . Then, Ewc = inf M =M ! Tr QM subject to: z " M z ≥ φ(x), ∀x ∈ Rn (6) and Pwc = inf Tr QM M &0 subject to: z " M z ≥ 1, ∀x ∈ S. (7) & A proof of Lemma 1 is given in the Appendix. Remark 2 Lemma 1 provides a result for computing worst-case expectations and probabilities. However in many cases of interest we shall need to impose constraints on these quantities in order to eventually optimize them with respect to some other design variables. In this respect, the following equivalence holds: supx∼(x̂,Γ) E {φ(x)} ≤ γ . " ∃M = M : Tr QM ≤ γ, z " M z ≥ φ(x), ∀x ∈ Rn . (8) (9) To verify this fact, consider first the ⇓ direction: If (8) holds, we let M be the solution that achieves the optimum in (6), and we have that (9) holds. On the other hand, if (9) holds for some M = M̄ , then supx∼(x̂,Γ) E {φ(x)} = inf Tr QM ≤ Tr QM̄ ≤ γ, which concludes proves the statement. By an analogous reasoning, we can verify that supx∼(x̂,Γ) Prob {x ∈ S} ≤ $ . ∃M = M " : M ( 0, Tr QM ≤ $, z " M z ≥ 1, ∀x ∈ S. & 4 To appear in: Systems and Control Letters (Elsevier) Worst-case expected residual minimization 3 In this section we focus on Problem 1 and provide an efficiently computable exact solution for the case p = 2, and efficiently computable upper and lower bounds on the solution for the case p = 1. Define . ψp (θ) = sup E {#L(θ)z#p }, . with z " = [δ " 1], (10) . r = [0 · · · 0 1/2]" ∈ Rq+1 , (11) δ∼(0,I) where L(θ) ∈ Rm,q+1 is an affine function of parameter θ, given in (3), (4). We have the following preliminary lemma. Lemma 2 For given θ ∈ Rn , the worst-case residual expectation ψp (θ) is given by ψp (θ) = inf M =M ! Tr M subject to: M − ru" L(θ) − L(θ)" ur" ( 0, ∀u ∈ Rm : #u#p∗ ≤ 1, where #u#p∗ is the dual "p norm. & Proof. From Lemma 1 we have that ψp (θ) = inf M =M ! Tr M subject to: z " M z ≥ #L(θ)z#p , ∀δ ∈ Rq . Since #L(θ)z#p = sup u" L(θ)z, 'u'p∗ ≤1 it follows that z " M z ≥ #L(θ)z#p holds for all δ if and only if z " M z ≥ u" L(θ)z, ∀δ ∈ Rq and ∀ u ∈ Rm : #u#p∗ ≤ 1. Now, since z " r = 1/2, we write u" L(θ)z = z " (ru" L(θ) + L(θ)" ur" )z, whereby the above condition is satisfied if and only if M − ru" L(θ) + L(θ)" ur" ( 0, ∀ u : #u#p∗ ≤ 1, ! which concludes the proof. We are now in position to state the following key theorem. Theorem 1 Let θ ∈ Rn be given, and let ψp (θ) be defined as in (10). Then, the following holds for the worst-case expected residuals in the "1 - and "2 -norm cases. 1. Case p = 1: Define m & . #& &Li (θ)" & , ψ 1 (θ) = 2 (12) i=1 where Li (θ)" denotes the i-th row of L(θ). Then, 2 ψ (θ) ≤ ψ1 (θ) ≤ ψ 1 (θ). π 1 5 (13) To appear in: Systems and Control Letters (Elsevier) 2. Case p = 2: ψ2 (θ) = ' Tr L(θ)" L(θ) = #L(θ)#F . (14) & Proof. (Case p = 1) The dual "1 norm is the "∞ norm, hence applying Lemma 2 we have ψ1 (θ) = inf M =M ! Tr M subject to: M − L(θ) ur" − ru" L(θ) ( 0, " (15) ∀u : #u#∞ ≤ 1. For ease of notation, we drop the dependence on θ in the following derivation. Note that " " " L ur + ru L = m # ui Ci , i=1 where . " Ci = rL" i + Li r = ( 0q 1 (δ)" L 2 i 1 (δ) L 2 i (1) Li ) (δ)" ∈ Rq+1,q+1 , (1) (δ)" " where L" Li ], with Li ∈ R1,q , and i is partitioned according to (4) as Li = [Li (1) (1) (δ) Li ∈ R. The characteristic polynomial of Ci is pi (s) = sq−1 (s2 − Li s − #Li #22 /4), hence (1) Ci has q − 1 null eigenvalues, and two non-zero eigenvalues at ηi,1 = (Li + #Li #2 )/2 > 0, (1) ηi,2 = (Li − #Li #2 )/2 < 0. Since Ci is rank two, the constraint in problem (15) takes the form (23) considered in Theorem 4 in the Appendix. Consider thus the following relaxation of problem (15): . ϕ= inf Tr M (16) M =M ! ,Xi =Xi! subject to: −Xi + Ci 1 0, − Xi − Ci 1 0, i = 1, . . . , m, *m i=1 Xi − M 1 0, where we clearly have ψ1 ≤ ϕ. The dual of problem (16) can be written as D ϕ = sup Λi ,Γi subject to: m # i=1 Tr ((Λi − Γi )Ci ) (17) Λi + Γi = Iq+1 , Γi ( 0, Λi ( 0, i = 1, . . . , m. Since the problem in (16) is convex and Slater conditions are satisfied, ϕ = ϕD . Next we show that ϕD equals ψ 1 given in (12). To this end, observe that (17) is decoupled in the Γi , Λi variables and, for each i, the subproblem amounts to determining sup0*Γi *I Tr (I − 2Γi )Ci . By diagonalizing Ci as Ci = Vi Θi Vi" , with Θi = diag(0, . . . , 0, ηi,1 , ηi,2 ), each subproblem is reformulated as sup0*Γ̃i *I Tr Ci −2Tr Θi Γ̃i , where it immediately follows that the optimal solution is Γ̃i = diag(0, . . . , 0, 0, 1), hence the supremum is (ηi,1 + ηi,2 ) − 2ηi,2 = ηi,1 − ηi,2 = |ηi,1 | + |ηi,2 | = #eig(Ci )#1 , where eig(·) denotes the*vector of the eigenvalues of m D " its argument. Now, we have #eig(Ci )#1 = #L" i #2 , then ϕ = i=1 #Li #2 , and by the first D conclusion in Theorem 4 in the Appendix, we have ψ 1 = ϕ = ϕ and ψ1 ≤ ψ 1 . For the lower bound on ψ1 in (13), assume that the problem in (17) is not feasible. Then, for M ( 0, we have that + ,M : Tr M = ϕD *n {M : Xi ( ±Ci , i=1 Xi 1 M } = ∅. 6 To appear in: Systems and Control Letters (Elsevier) This last emptiness statement, coupled with the fact that, for i = 1, . . . , n, Ci is of rank two, implies, by the second conclusion in Theorem 4, that + ,M : Tr M = ϕD * {M : M ( ni=1 ui Ci , ∀u : |ui | ≤ π/2} = ∅ and ! "ϕD M̃ : Tr M̃ = π/2 ! " *n M̃ : M̃ ( i=1 ũi Ci , ∀ũ : |ũi | ≤ 1 = ∅ Consequently, we have ψ1 ≥ ϕD π/2 = ψ1 , π/2 which concludes the proof of the p = 1 case. (Case p = 2) The dual "2 norm is the "2 norm itself, hence applying Lemma 2 we have ψ2 = inf M =M ! Tr M subject to: M − ru" L − L" ur" ( 0, ∀u : #u#2 ≤ 1. Applying the LMI robustness lemma (Lemma 3.1 of [9]), we have that the previous semiinfinite problem is equivalent to the following SDP $ % M − τ rr" L" ψ2 (θ) = inf Tr M subject to: ( 0. L τ Im M =M ! ,τ >0 By the Schur complement rule, the latter constraint is equivalent to τ > 0 and M ( 1 (L" L) + τ rr" . Thus, the infimum of Tr M is achieved for M = τ1 (L" L) + τ rr" and, since τ √ Tr L" L. From rr" = diag(0q , 1/4), the infimum of Tr M over τ > 0 is achieved for τ = 2 √ this, it follows that ψ2 = Tr L" L, thus concluding the proof. ! Starting from the results in Theorem 1, it is easy to observe that we can further minimize the residuals over the parameter θ, in order to find a solution to Problem 1. Convexity of the ensuing minimization problem is a consequence of the fact that L(θ) is an affine function of θ. This is formalized in the following corollary, whose simple proof is omitted. Corollary 1 (Worst-case expected residual minimization) Let . ψp∗ = minn sup E {#L(θ)z#p }, θ∈R . z " = [δ " 1]. δ∼(0,I) For p = 1, it holds that 2 ∗ ∗ ψ 1 ≤ ψ1∗ ≤ ψ 1 , π ∗ where ψ 1 is computed by solving the following second-order-cone (SOCP) program: ∗ ψ 1 = minn θ∈R m # i=1 #Li (θ)" #2 . For p = 2, it holds that ψ2∗ = minn #L(θ)#F , θ∈R where a minimizer for this problem can be computed via convex quadratic programming, by minimizing Tr L" (θ)L(θ). & 7 To appear in: Systems and Control Letters (Elsevier) 3 Notice that in the specific case of δ ∼ (0, I) we have that ψ22 = Tr L" (θ)L(θ) = Remark *q 2 i=0 #Ai θ − bi #2 , hence the minimizer can in this case be determined by standard LeastSquares solution method. Interestingly, this solution coincides with the solution of the expected squared "2 -norm minimization problem discussed for instance in [4, 10]. This might not be obvious, since in general E {# · #2 } 3= (E {# · #})2 . & 4 4.1 Guaranteed residual-at-risk minimization The "2 -norm case Assume first θ ∈ Rn is fixed, and consider the problem of computing Pwc2 (θ) = sup Prob {δ : #L(θ)z#2 ≥ γ} = sup Prob {δ : #L(θ)z#22 ≥ γ 2 }, δ∼(0,I) δ∼(0,I) . where z " = [δ " 1]. By Lemma 1, this probability corresponds to the optimal value of the optimization problem Pwc2 (θ) = inf Tr M M &0 subject to: z " M z ≥ 1, ∀δ : #L(θ)z#22 ≥ γ 2 , where the constraint can be written equivalently as . / z " (M − diag(0q , 1)) z ≥ 0, ∀δ : z " L(θ)" L(θ) − diag(0q , γ 2 ) z ≥ 0. Applying the lossless S-procedure, the condition above is in turn equivalent to the existence . / of τ ≥ 0 such that (M − diag(0q , 1)) ( τ L(θ)" L(θ) − diag(0q , γ 2 ) , therefore we obtain Pwc2 (θ) = inf M &0,τ >0 Tr M subject to: M ( τ L(θ)" L(θ) + diag(0q , 1 − τ γ 2 ), where the latter expression can be further elaborated using the Schur complement formula into $ % M − diag(0q , 1 − τ γ 2 ) τ L(θ)" ( 0. (18) τ L(θ) τ Im We now notice, by the reasoning in Remark 2, that the condition Pwc2 (θ) ≤ $ with $ ∈ (0, 1) is equivalent to the conditions ∃τ ≥ 0, M ( 0 such that: Tr M ≤ $, and (18) holds. A parameter θ that minimizes the residual-at-risk level γ while satisfying the condition Pwc2 (θ) ≤ $ can thus be computed by solving a sequence of convex semidefinite optimization problems parameterized in τ , as formalized in the next theorem. Theorem 2 ("2 residual-at-risk estimation) A solution of Problem 2 in the "2 case can be found by solving the following optimization problem: inf γ 2, inf τ >0 M &0,θ∈Rn ,γ 2 ∈Rn subject to: $ Tr M ≤ $ % M − diag(0q , 1 − τ γ 2 ) τ L(θ)" ( 0. τ L(θ) τ Im 8 (19) & To appear in: Systems and Control Letters (Elsevier) Remark 4 The optimization problem in Theorem 2 is not jointly convex in all variables, due to the presence of product terms τ γ 2 , τ L(θ). However, the problem is convex in M, θ, γ 2 for each fixed τ > 0, whence an optimal solution can be computed via a line search over τ > 0, where each step requires solving an SDP in M , θ, and γ 2 . & 4.2 The "1 -norm case We next consider the problem of determining θ ∈ Rn such that the residual-at-risk level γ is minimized while guaranteeing that Pwc1 (θ) ≤ $, where Pwc1 (θ) is the worst-case "1 -norm residual tail probability Pwc1 (θ) = sup Prob {δ : #L(θ)z#1 ≥ γ}, δ∼(0,I) and $ ∈ (0, 1) is the a-priori fixed risk level. To this end, define . D = {D ∈ Rm,m : D diagonal, D ' 0} and consider the following proposition (whose statement may be easily proven by taking the gradient with respect to D and setting it to zero). Proposition 1 For any v ∈ Rm , it holds that 1 m 0 2 # . / vi 1 1 + di = inf v " D−1 v + Tr D , #v#1 = inf 2 D∈D i=1 di 2 D∈D where di is the i-th diagonal entry of D. (20) & The following key theorem holds. Theorem 3 ("1 residual-at-risk estimation) Consider the following optimization problem: inf inf γ τ >0 M &0,D∈D,θ∈Rn ,γ≥0 subject to: $ (21) Tr M ≤ $ % M − (1 − 2τ γ + Tr D)J τ L(θ)" ( 0, τ L(θ) D . with J = diag(0q , 1). The optimal value of this program provides an upper bound for Problem 2 in the "1 case, that is an upper bound on the minimum level γ for which there exist θ such that Pwc1 (θ) ≤ $. & Remark 5 Similar to the problem (19) in Theorem 2, the above optimization problem is non convex, due to product terms between τ and γ, θ. The problem can however be easily solved numerically via a line search over the scalar τ > 0, where each step of the line search involves the solution of an SDP problem in the variables M , D, θ, and γ. & Proof. Define . S = {δ : #L(θ)z#1 ≥ γ} . S(D) = {δ : z " L(θ)" D−1 L(θ)z + Tr D ≥ 2γ, D ∈ D}. 9 To appear in: Systems and Control Letters (Elsevier) For ease of notation we drop the dependence on θ in the following derivation. Using (20) we have that, for any D ∈ D, 2#Lz#1 ≤ z " L" D−1 Lz + Tr D, hence δ ∈ S implies δ ∈ S(D), thus S ⊆ S(D), for any D ∈ D. This in turn implies that Prob {δ ∈ S} ≤ Prob {δ ∈ S(D)} ≤ sup Prob {δ ∈ S(D)} δ∼(0,I) for any probability measure and any D ∈ D, and therefore . Pwc1 = sup Prob {δ ∈ S} ≤ inf sup Prob {δ ∈ S(D)} = P̄wc1 . D∈D δ∼(0,I) δ∼(0,I) . Note that, for fixed D ∈ D, we can compute Pwc1 (D) = supδ∼(0,I) Prob {δ ∈ S(D)} from its equivalent dual: Pwc1 (D) = = inf Tr M : z " M z ≥ 1, ∀δ ∈ S(D) M &0 inf Tr M : z " M z ≥ 1, ∀δ : z " L" D−1 Lz + Tr D ≥ 2γ M &0 [applying the lossless S-procedure] = inf Tr M : M ( τ L" D−1 L + (1 − 2τ γ + τ Tr D)J, M &0,τ >0 where J = diag(0q , 1). Hence, P̄wc1 is obtained by minimizing Pwc1 (D) over D ∈ D, which results in P̄wc1 = inf M &0,τ >0,D∈D Tr M : M ( τ L" D−1 L + (1 − 2τ γ + τ Tr D)J [by change of variable τ D → D (whence τ Tr D → Tr D, τ D−1 → τ 2 D−1 )] = inf Tr M : M ( τ 2 L" D−1 L + (1 − 2τ γ + Tr D)J M &0,τ >0,D∈D $ % M − (1 − 2τ γ + Tr D)J τ L" = inf Tr M : ( 0. τL D M &0,τ >0,D∈D Now, from the reasoning in Remark 2, we have that (re-introducing the dependence on θ in the notation) P̄wc1 (θ) ≤ $ if and only if there exist M ( 0, τ > 0 and D ∈ D such that Tr M ≤ $ and $ % M − (1 − 2τ γ + Tr D)J τ L(θ)" ( 0. (22) τ L(θ) D Notice that, since L(θ) is affine in θ , condition (22) is an LMI in M, D, θ, γ, for fixed τ . We can thus minimize the residual level γ subject to the condition P̄wc1 (θ) ≤ $ (which implies Pwc1 (θ) ≤ $), and this results in the statement of the theorem. ! 5 5.1 Numerical examples Worst-case expected residual minimization As a first example, we use data from a numerical test appeared in [4]. Let A(δ) = A0 + 3 # δi Ai , i=1 10 bT = 2 3 0 2 1 3 , To appear in: Systems and Control Letters (Elsevier) 3 1 4 0 0 0 0 0 1 0 0 0 0 1 1 0 0 0 0 1 0 0 0 0 with A0 = −2 5 3 A1 = 0 0 0 , A2 = 0 0 0 , A3 = 1 0 0 , 1 4 5.2 0 0 1 0 0 0 0 0 0 and let δi be independent random perturbations with zero mean and standard deviations σ1 = 0.067, σ2 = 0.1, σ3 = 0.2. The standard "2 and "1 solutions (obtained neglecting the uncertainty terms, i.e. setting A(δ) = A0 ) result to be −10 −11.8235 θnom2 = −9.728 , θnom1 = −11.5882 , 9.983 11.7647 with nominal residuals of 1.7838 and 1.8235, respectively. Applying Theorem 1, the minimal worst-case expected "2 residual resulted to be ψ2∗ = 2.164, whereas the minimal upper bound on worst-case expected "1 residual resulted to be ψ̄1∗ = 4.097. The corresponding parameter estimates are −2.3504 −2.8337 θewc2 = −2.0747 , θewc1 = −2.5252 . 2.4800 2.9047 We next analyzed numerically how the worst-case expected residuals increase with the level of perturbation. To this end, we consider the previous data with standard deviations on the perturbation depending on a parameter ρ ≥ 0: σ1 = ρ · 0.067, σ2 = ρ · 0.1, σ3 = ρ · 0.2. A plot of the worst-case expected residuals as a function of ρ is shown in Figure 1. We observe that both "1 and "2 expected residuals tend to a constant value for large ρ. 4.5 ψ∗ 1 4 3.5 3 2.5 ψ∗ 2 2 1.5 0 0.5 1 1.5 2 2.5 ρ Figure 1: Plot of ψ2∗ (solid) and ψ̄1∗ (dashed) as a function of perturbation level ρ. 11 3 To appear in: Systems and Control Letters (Elsevier) Guaranteed residual at risk minimization 5.2 As a first example of guaranteed residual at risk minimization, we consider again the variable perturbation level problem of the previous section. Here, we fix the risk level to $ = 0.1 and solve repeatedly problems (19) and (21) for increasing values of ρ. A plot of the resulting optimal residuals at risk as a function of ρ is shown in Figure 2. These residuals grow with the covariance level ρ, as it might be expected since increasing the covariance increases the tails of the residual distribution. 4.32 ||.|| 1 residual @ risk 0.1 ||.|| 2 residual @ risk 0.1 2.38 2.36 2.34 4.3 4.28 4.26 2.32 4.24 2.3 4.22 2.28 4.2 4.18 0 0.5 1 1.5 2 2.5 ρ 3 0 0.5 1 1.5 2 2.5 ρ 3 Figure 2: Worst-case "2 and "1 residuals at risk as a function of perturbation level ρ. As a further example, we consider a system identification problem where one seeks to estimate the impulse response of a discrete-time linear FIR system from its input/output measurements. Let the system be described by the convolution yk = N # j=1 hj · uk−τ +1 , k = 1, 2, . . . , where ui is the input, yi is the output, and h = [h1 · · · hN ] is the impulse response to be . estimated. If N input/output measurements are collected in vectors u" = [u1 u2 · · · uN ], . y " = [y1 y2 · · · yN ], then the impulse response h can be computed by solving the system of linear equations U h = y, where U is a lower-triangular Toeplitz matrix having u in the first column. * In practice, however, *Nboth u and y might be affected by errors, that is U (δu ) = U + N δ U , y(δ ) = y + y i=1 ui i i=1 δyi ei where ei is the i-th column of the identity N matrix in R , and Ui is a lower-triangular Toeplitz matrix having ei in the first column. These uncertain data are easily re-written in the form (1) by setting θ = h, q = 2N , A0 = U , b0 = y, and, for i = 1, . . . , 2N , : : : δui , if i ≤ N 0, if i ≤ N Ui , if i ≤ N , δi = Ai = , bi = δy,i−N , otherwise ei−N , otherwise 0N , otherwise 12 To appear in: Systems and Control Letters (Elsevier) We considered the same data used in a similar example in [8], that is u" = [1 2 3], y " = [4 5 6], and assumed that the input/output measurements are affected by i.i.d. errors with zero mean and unit variance. The nominal system U h = y has minimum-norm solution hnom = [4 − 3 0]" . First, we solved the "2 residual-at-risk problem in (19), setting risk level $ = 0.1. The optimal solution was achieved for τ = 0.013 and yielded an optimal worst-case residual at risk γ = 11.09, with corresponding parameter h(2 ,rar = [0.7555 0.4293 0.1236]" . This means that, no matter what the distribution on the uncertainty is (as long has it has the assumed mean and covariance), the probability of having a residual larger than γ = 11.09 is smaller than $ = 0.1. We next performed an a-posteriori Monte-Carlo test of this solution against the nominal one hnom , for two specific noise distributions, namely the uniform and the Normal distribution. The empirical cumulative distribution functions of the resulting norm residuals are shown in Figure 3. Some relevant statistics are also shown in Table 1. Prob{||L(θ)z|| < γ} (a) Uniform noise distribution (b) Normal noise distribution 1 1 0.9 0.9 0.8 0.8 0.7 0.7 0.6 0.6 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 0 0 2 4 6 8 10 12 14 16 18 γ 20 0 5 10 15 20 25 γ 30 Figure 3: Empirical "2 -norm residual cumulative distribution for the h(2 ,rar solution (solid) and the nominal hnom solution (dashed). Simulation in case (a) assumes uniform noise distribution; case (b) assumes Normal distribution. min max mean median std rar@$ = 0.1 Uniform hnom h(2 ,rar 0.19 0.49 19.40 10.79 7.75 5.52 7.50 5.53 3.02 1.44 11.85 7.39 Normal hnom h(2 ,rar 0.19 0.15 28.98 11.75 7.50 5.51 7.01 5.49 3.58 1.43 12.39 7.36 Table 1: "2 -norm residual statistics from a-posteriori test on nominal and residual-at-risk solutions, with uniform and Normal distribution on the noise. We next solved the same problem using an "1 -norm residual criterion. In this case, solution of problem (21) yielded τ = 0.082 and an optimal upper bound on "1 residual at risk 13 To appear in: Systems and Control Letters (Elsevier) Prob{||L(θ)z||1< γ} (a) Uniform noise distribution (b) Normal noise distribution 1 1 0.9 0.9 0.8 0.8 0.7 0.7 0.6 0.6 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 0 5 10 15 20 25 30 γ 35 0 0 5 10 15 20 25 30 35 40 45 γ 50 Figure 4: Empirical "1 -norm residual cumulative distribution for the h(1 ,rar solution (solid) and the nominal hnom solution (dashed). Simulation in case (a) assumes uniform noise distribution; case (b) assumes Normal distribution. γ = 19.2, with corresponding parameter h(1 ,rar = [0.7690 0.4464 0.1344]" . An a-posteriori Monte-Carlo test then produced the residual distributions shown in Figure 4. Residual statistics are reported in Table 2. min max mean median std rar@$ = 0.1 Uniform hnom h(1 ,rar 0.38 0.81 32.12 18.39 11.89 8.95 11.34 8.95 5.01 2.61 18.66 12.37 Normal hnom h(1 ,rar 0.11 0.57 47.93 19.63 11.42 8.94 10.52 8.93 5.76 2.56 19.22 12.29 Table 2: "1 -norm residual statistics from a-posteriori test on nominal and residual-at-risk solutions, with uniform and Normal distribution on the noise. 6 Conclusions In this paper we discussed two criteria for linear parameter estimation in presence of random uncertain data, under both "2 and "1 norm residuals. The first criterion is a worst-case residual expectation and leads to exact and efficiently computable solutions for the "2 norm case. For the "1 norm, we can efficiently compute upper and lower bounds on the optimal solution, by means of convex second order cone programming. The second criterion considered in the paper is the worst-case residual for a given risk level $. With this criterion, an exact solution for the "2 norm case can be computed by a solving a sequence of 14 To appear in: Systems and Control Letters (Elsevier) semi-definite programs, and an analogous computational effort is required for computing an upper bound on the optimal solution in the "1 norm case. The estimation setup proposed in the paper is “distribution free,” in the sense that only information about the mean and covariance of the random uncertainty need be available to the user: the results are guaranteed irrespective of the actual shape of uncertainty distribution. Appendix A Proof of Lemma 1 We start by recalling a preliminary result. Let fi : Rn → R, i = 1, . . . , m, be functions whose expectations qi are given and finite, and consider the following problem (P ) and its dual (D): (P ) : . Z P = supx E {φ(x)} subject to: E {fi (x)} = qi , i = 0, 1, . . . , m; (D) : . Z D = inf y E {y " f (x)} = inf y " q subject to: y " f (x) ≥ φ(x), ∀x ∈ Rn , where f0 (x) = 1 and q0 = 1 correspond to the implied probability-mass constraint, and the infimum and the supremum are taken with respect to all probability distributions satisfying the specified moment constraints. Then, strong duality holds: Z P = Z D , hence supx E {φ(x)} can be computed by solving the dual problem (D); see for instance Section 16.4 of [2]. Now, the primal problem that we wish to solve in Lemma 1 for computing Ewc is (P ) : Z P = Ewc = supx E {φ(x)} subject to: E {x} = x̂ E {xx" } = Γ + x̂x̂" , where Γ ' 0 is the covariance matrix of x, and the functions fi are xk , k = 1, . . . , n, and xk xj , 1 ≤ k ≤ j ≤ n. Then, the dual problem is (D) : Z D = inf y0 ∈R,y∈Rn ,Y =Y ! ∈Rn,n y0 + y " x̂ + Tr (Γ + x̂x̂" )Y subject to: y0 + y " x + Tr xx" Y ≥ φ(x), ∀x ∈ Rn , where the dual variable y0 is associated with the implicit probability-mass constraint. Defining % $ % $ % $ 1 y Γ + x̂x̂" x̂ x Y 2 , Q= , z= , M= 1 " " y y0 x̂ 1 1 2 this latter problem writes (D) : Z D = inf M =M ! ∈Rn+1 Tr QM subject to: z " M z ≥ φ(x), ∀x ∈ Rn , which is (6), thus concluding the first part of the proof. 15 To appear in: Systems and Control Letters (Elsevier) The result in (7) can then be obtained by specializing (6) to the case when φ(x) = IS (x), where IS is the indicator function of set S, since Prob {x ∈ S} = E {IS (x)}. We thus have that Z P = Z D for Z P = Pwc = supx Prob {x ∈ S} subject to: E {x} = x̂, E {xx" } = Γ + x̂x̂" , (D) : Z D = inf M =M ! ∈Rn+1 Tr QM subject to: z " M z ≥ IS (x), ∀x ∈ Rn . (P ) : The constraint z " M z ≥ IS (x) ∀x ∈ Rn can be rewritten as z " M z ≥ 1 ∀x ∈ S, z " M z ≥ 0 ∀x ∈ Rn , and this latter constraint is equivalent to requiring M ( 0, which explains (7) and concludes the proof. ! B Matrix cube theorem Theorem 4 (Matrix cube relaxation; [1]) Let B 0 , B 1 , . . . , B L ∈ Rn×n be symmetric and B 1 , . . . , B L be of rank two. Let the problem Pρ be defined as: 0 Pρ : Is B + L # i=1 ui B i ( 0, ∀u : #u#∞ ≤ ρ ? (23) and the problem Prelax be defined as: Prelax : Do there exist symmetric matrices X1 , . . . , XL ∈ Rn×n satisfying Xi ( ±ρB i , i = 1, . . . , L, *L 0 i=1 Xi 1 B ? Then, the following statements hold: 1. If Prelax is feasible, then Pρ is feasible. 2. If Prelax is not feasible, then P π2 ρ is not feasible. " References [1] A. Ben-Tal and A. Nemirovski. On tractable approximations of uncertain linear matrix inequalities affected by interval uncertainty. SIAM Journal on Optimization, 12(3):811–833, 2002. [2] D. Bertsimas and J. Sethuraman. Moment problems and semidefinite optimization. In Handbook of semidefinite programming, pages 469–509. Kluwer Acad. Publ., Boston, MA, 2000. [3] S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge University Press, New York, NY, 2004. ISBN: 0-521-83378-7. [4] G. Calafiore and F. Dabbene. Near optimal solutions to least-squares problems with stochastic uncertainty. Systems and Control Letters, 54(12):1219–1232, December 2005. 16 To appear in: Systems and Control Letters (Elsevier) [5] C. Dossal. A necessary and sufficient condition for exact recovery by "1 minimization. http://hal.archives-ouvertes.fr/hal-00164738/fr/, 2008. [6] D. Duffie and J. Pan. An overview of value at risk. Journal of Derivatives, pages 7–49, Spring 1997. [7] M.P. Friedlander and P. Tseng. Exact regularization of convex programs. SIAM J. on Optimization, 18(4):1326–1350, 2007. [8] L. El Ghaoui and H. Lebret. Robust solutions to least-squares problems with uncertain data. SIAM J. on Matrix Analysis and Applications, 18(4):1035–1064, 1997. [9] L. El Ghaoui, F. Oustry, and H. Lebret. Robust solutions to uncertain semidefinite programs. SIAM J. Optimization, 9(1):33–52, 1998. [10] H.A. Hindi and S.P. Boyd. Robust solutions to l1 , l2 , and l∞ uncertain linear approximation problems using convex optimization. In Proc. American Control Conf., volume 6, pages 3487–3491, 1998. [11] R.T. Rockafellar and S. Uryasev. Conditional value-at-risk for general loss distributions. Journal of banking and Finance, 26:1443–1471, 2002. [12] A.H. Sayed, V.H. Nascimento, and F.A.M. Cipparrone. A regularized robust design criterion for uncertain data. SIAM J. on Matrix Analysis and Applications, 23(4):1120–1142, 2002. 17